A0

A0

The programming language built for AI, not for people. a0lang.com · Docs

A0 is a compact, exactly specified language that models write and edit through revision-checked structured edits. A model reads only what an edit touches, writes only the changed lines, and nothing invalid lands. One program compiles to native machine code (A0's own AArch64 code generator, or C), the browser (wasm32), JavaScript, the JVM, .NET, Metal GPU kernels, and clocked SystemVerilog, and every target is verified against one oracle.

Measured on the repository's benchmarks (Apple M3, 8 cores, results/*.json). The machine was not quiet: results/exec-benchmark.json records a 1/5/15-minute load average of 6.4-8.0 for the main and arm64 runs and 32-40 for the JavaScript remeasurement; a quiet-machine rerun is pending:

The site a0lang.com is itself two A0 programs (site/page.a0, site/docs.a0).

See DESIGN.md for intent and semantics, MODEL_GUIDE.txt for the AI-facing language instructions, STATUS.md for the current results, loss ledger, known limits, and next actions (session history in docs/history/), and results/ for machine-readable evidence.

Install

A0 is a single self-contained binary: no Node, no Bun.

# macOS / Linux: Homebrew (this repository is the tap; no separate tap repository)brew tap Joe-Simo/a0 https://github.com/Joe-Simo/a0brew install a0
# macOS / Linux: script (detects OS/arch, verifies SHA-256, installs to ~/.local/bin)curl -fsSL https://raw.githubusercontent.com/Joe-Simo/a0/main/install.sh | sh

# Windows (installs to %LOCALAPPDATA%\Programs\a0 and adds it to your user PATH)irm https://raw.githubusercontent.com/Joe-Simo/a0/main/install.ps1 | iex

Both scripts accept A0_VERSION=v0.8.16 to pin a release and A0_INSTALL_DIR to change the destination. Or download a binary from the latest release by hand and check it against checksums.txt. a0 --version prints the release and compiler version. Then:

printf 'fn sq u32 -> u32\na mul p0 p0\nret a\nend\n' > sq.a0a0 run sq.a0 sq 12            # 144a0 emit arm64 sq.a0           # A0's own machine code; or c, js, java, sva0 check sq.a0                # diagnostics with codes; --json for fields, --fix applies the exact fixesa0 explain A0102              # what a diagnostic means, with a failing and a fixed example

Binaries: a0-darwin-arm64, a0-darwin-x64, a0-linux-x64, a0-linux-arm64, a0-windows-x64.exe. A C compiler (clang or gcc) is needed only to link native output on your machine.

MCP server

a0 mcp <file-or-dir> serves A0 to AI agents over stdio (Model Context Protocol), so they edit through tools instead of text files: a0_open (function view with a handle), a0_program (program handle, optionally scoped to a target), a0_apply (edit under a handle; returns the new view or a diagnostic with code (class), id (A0nnnn), message, expected/actual, fix and applicability; the reply fix all applies every exact fix of the last rejected edit, atomically), a0_check, a0_run (reference interpreter, fuel-bounded), a0_emit (any target), a0_save (only after a successful apply). Paths are confined to the launch root (symlink escapes and use escapes rejected); no shell is run; output is bounded.

Language server

a0 lsp [root] is a Language Server Protocol server over stdio (built on vscode-languageserver; --stdio is accepted and ignored). Point any LSP client at it for .a0 files: diagnostics on open and change from the linker and checker (the diagnostic code is the class, data carries the table code id, the fix and its applicability, and an exact fix is a quick fix), hover (function signatures, op docs), go-to-definition across use files, document symbols, completion of ops and in-scope functions, and formatting through the canonical printer. Files are confined to root (default: the working directory) exactly as in the MCP server; a document outside it gets one limit diagnostic and nothing else.

Use with AI agents

Put a0 on your PATH first (see Install). Every entry below runs the same local stdio server, a0 mcp <dir>, with no hosting and no account. Ready-to-copy configs are in integrations/. The Agent Skill in skills/a0/ (install with npx skills add Joe-Simo/a0) teaches the language and loads the primer and the edit protocol only when they are needed.

Replace . with the folder the server may read and write when your agent does not start servers in the project folder.

Project rules, hooks and CI

• a0 init [dir] writes AGENTS.md (read by Codex, Cursor, Copilot, Jules and others) with the MCP tool workflow, the edit protocol and the primer path (.a0/MODEL_GUIDE.txt, copied in when available, otherwise the GitHub URL), plus @AGENTS.md pointers in CLAUDE.md and GEMINI.md, glob rules for Cursor (.cursor/rules/a0.mdc), Windsurf (.windsurf/rules/a0.md) and Copilot (.github/instructions/a0.instructions.md), and a Gemini CLI AfterTool hook (.gemini/settings.json). Existing instruction files get the A0 section appended once; other existing files are kept.
• a0 hook is the after-edit check: it reads a hook's tool-call JSON on stdin and, when an edited *.a0 file fails a0 check, prints a blocking response whose reason is the exact diagnostic with its fix:.
• Claude Code plugin (plugin/): a PostToolUse hook on Edit|Write|MultiEdit (plugin/hooks/hooks.json, runs a0 hook; set A0_BIN if a0 is not on PATH) and the slash commands /a0-check, /a0-emit, /a0-run (plugin/commands/).
• CI: uses: Joe-Simo/a0@main (root action.yml, input version, default latest) downloads the release binary for the runner and checks every tracked .a0 file. pre-commit: repo: https://github.com/Joe-Simo/a0, hook id a0-check (needs a0 on PATH).
• a0 check accepts several files and exits 1 if any fails, printing <file>: error: <code>: ... fix: ... per failure.

Privacy

The a0 MCP server and the .mcpb bundles run entirely on your machine. They collect no data, make no network requests, and read or write only inside the folder you give them.

Contributing to the compiler

See CONTRIBUTING.md for the layout, the edit protocol, how to add a backend or a language to the experiments, and the measurement rules.

The compiler is being rewritten in A0 (see compiler/ and DESIGN.md section 7a). Until that lands, the compiler itself is TypeScript, and working on it needs Bun or Node 22+. Users of A0 never need this: the released a0 binary is self-contained.

bun installbun run lint        # Biomebun run typecheck   # tsc --noEmitbun run test        # node --test on dist/testbun run verify      # cross-toolchain differential execution -> results/verification.jsonbun run hw          # Icarus simulation + Yosys synthesis  -> results/hardware.jsonbun run bench       # in-process timings + byte fixture     -> results/benchmark.jsonbun run tokens      # tokenizer probe (js-tiktoken)         -> results/tokens.jsonbun run exec-bench  # emitted vs 45+ hand-written baselines  -> results/exec-benchmark.jsonbun run app         # Life application acceptance, 7 targets -> results/app.jsonbun run site        # browser demo (wasm + DOM adapter)     -> site/dist/bun run gpu         # Metal GPU execution of the corpus     -> results/gpu.jsonbun run dotnet      # C# / .NET execution of the corpus     -> results/dotnet.json

CLI (after bun run build):

node dist/src/cli.js check examples/kernels.a0node dist/src/cli.js run examples/kernels.a0 affine 10 3 7      # 37node dist/src/cli.js emit js|c|java|sv examples/kernels.a0 [out]node dist/src/cli.js wasm examples/kernels.a0 kernels.wasm       # needs clang + wasm-ldnode dist/src/cli.js revision examples/kernels.a0 affinenode dist/src/cli.js patch examples/kernels.a0 edit.patch [out.a0]

Layout: src/core.ts grammar/validation/interpreter, src/edit.ts revisions and edit sessions, src/optimize.ts, src/backends.ts JS/C/Java/SystemVerilog emission and emission cache, src/toolchain.ts installed-tool integration, src/cli.ts, test/, tools/ (corpus + oracle, verify, hw-verify, bench, token-bench).

Errors built for agents

Every diagnostic is one row of src/diagnostics.ts: a stable code (A0nnnn), a coarse class (parse, type, structure, limit, edit, patch, revision, handle, runtime, cli, unchanged), a message template, the fix, and a0 explain text with one failing and one fixed example. A diagnostic reaches an agent as {code, id, message, line, expected, actual, fix, applicability, edits} from a0 check --json, the edit protocol, the MCP server and the LSP. An unknown name (op, function, node, fold body, loop predicate, type) gets a did-you-mean by TypeScript's spelling rule. A fix is exact (safe to apply blindly, with edits that do it) or maybe. In the edit protocol the reply fix all applies every exact fix of the last rejected reply and validates the result before it commits. test/diagnostics.test.ts runs every explain example, and corpus/reject/ holds rejected programs and replies with the diagnostic each must raise and the program its fix produces (node dist/tools/reject-corpus.js [--bless] [--coverage]).